View-Invariant Human Activity Recognition using Topic Model on Combined ORB-OF Feature

In this paper, we introduce a human activity recognition method that is able to adaptively handle viewpoint variations. For this work, we also propose a combined ORB-OF feature which is a combination of ORB (oriented FAST and rotated BRIEF) binary feature and optical flow (OF) motion feature; as well as a binary vector clustering algorithm to organize and normalize ORB binary descriptor vectors into clusters. In addition, our proposed method is robust to partial occlusion, subject scale and activity speed variations, as well as less sensitive to lighting condition and actor appearance. We validate the proposed method by conducting experiments using a self-recorded dataset.